/ho take the
kinds of the
d behaviors
are attended
sec. While
n, the head
, supporting
d as motion
action of the
mbining the
le key input,
ge the facial
behaviors.
umber of
persons
verlapped)
1.91
31
1. 22
JNS
sar that the
her than the
f presuming
ion. In our
ured by the
=
=.
—
processor
tecting
In general, it is difficult to extract the operator's head by
image processing in the practical scene. We captured
the movement of both pupils instead of the head motion,
because the pupils move together with the head and can
be easily extracted. Figure 2 illustrates the setup for the
recognition of the head motions. A monochrome CCD
camera (XC-75 : Sony Co.) is set between the CRT
display and the keyboard of the personal computer for the
word processor. The camera is turned upward and takes
the operators head. Even in the case that the operator's
face turns toward the keyboard, both pupils can be taken
by such the camera position and attitude. To extract the
pupils in distinction from the irises, the images are taken
in near infrared band. The high-pass (visible transmitting,
infrared absorbing) filter of the camera is demounted and
a low-pass (infrared transmitting, visible absorbing) filter
(IR-85: HOYA Co.) is attached to the camera. IR-LEDs
(infrared light emitting diodes: AN304: Stanley Co.) of
[Take facial image |
[ Calculate threshold level : th |
I
I
Binarization
1
| Dilation and contraction |
| Extract shape features |
| Capture candidates for pupils |
Number of
candidates for pupils
= 2
th > th -1
| Calculate mutual positional relations |
|
|. Select candidates for pupil pair |
Number of
candidtes for pupil pair
th => th -1
| Selecta pupil pair |
End
Figure3 Flow chart for extracting both pupils.
which the spectral emissivity peaks at a wavelength of
about 950 nm are used as light sources. The light
sources are attached around the personal computer.
Image processing system consists of an image
processing board (GPB-1 : Sharp Semiconductor Co.)
with an image processing library, a personal computer
(DESKPRO 4/33i : COMPAQ Co.) and a C language
compiler (MS-C/C++: Microsoft Co.). The ranges of
change in the size and the form of both pupils, the mutual
distance and the inclination of the straight line which ties
both pupils were measured for all sorts of the attitude of
some persons. The ranges with some margins are
defined as the conditions for extracting both pupils.
Figure 3 shows the flow chart for extracting both pupils.
First of all, binarization of the input image is done at the
threshold level based on the gray levels of the 50 sampling
points in the image. Next, unevenness of objects and the
number of holes and breaks in the objects are reduced by
doing dilation and contraction. Then, labeling is done to
the objects in the image. Candidates for pupils are
captured on the basis of their shape features: their
peripheral length, their circleness. If the number of the
candidates for pupils are 2 or more, then a suitable pair for
both pupils is extracted on the basis of the mutual
positional relations: Euclid distance between the objects,
the inclination of the straight line which ties the objects.
If the number of candidates for pupils are less than 2, or if
the number of candidates for pupil pair isn't 1, the
binarization is done again by subtracting 1 from the
threshold level. In order to keep extracting both pupils in
real time, the regions of interest on and after the second
frame are limited around both pupils by the following
method.: Let (r,, r,) and (, , /) be the coordinates of the
right pupil and the left pupil, respectively (see Figure 4).
Then the rectangular region which is defined by the
following inequalities is selected as the region of interest
on next frame.
pel sxal chr (1)
2 2
min(;,l,) += <y smax(r, J) + Lox (2)
where |, > r,
X Region of interest
/ /
y (Fx,Ty) (ly)
/
Image plane
Figure 4 The region of interest.
437
